Re: [RESEND PATCH] powerpc64/bpf: fix JIT code size calculation of bpf trampoline

[Hari Bathini <hbathini@linux.ibm.com>]

Hi Naveen,

Thanks for the review.

On 03/04/25 9:15 pm, Naveen N Rao wrote:
> On Wed, Mar 26, 2025 at 08:04:22PM +0530, Hari Bathini wrote:
>> The JIT compile of ldimm instructions can be anywhere between 1-5
>> instructions long depending on the value being loaded.
>>
>> arch_bpf_trampoline_size() provides JIT size of the BPF trampoline
>> before the buffer for JIT'ing it is allocated. BPF trampoline JIT
>> code has ldimm instructions that need to load the value of pointer
>> to struct bpf_tramp_image. But this pointer value is not same while
>> calling arch_bpf_trampoline_size() & arch_prepare_bpf_trampoline().
>> So, the size arrived at using arch_bpf_trampoline_size() can vary
>> from the size needed in arch_prepare_bpf_trampoline(). When the
>> number of ldimm instructions emitted in arch_bpf_trampoline_size()
>> is less than the number of ldimm instructions emitted during the
>> actual JIT compile of trampoline, the below warning is produced:
>>
>>    WARNING: CPU: 8 PID: 204190 at arch/powerpc/net/bpf_jit_comp.c:981 __arch_prepare_bpf_trampoline.isra.0+0xd2c/0xdcc
>>
>> which is:
>>
>>    /* Make sure the trampoline generation logic doesn't overflow */
>>    if (image && WARN_ON_ONCE(&image[ctx->idx] >
>> 			(u32 *)rw_image_end - BPF_INSN_SAFETY)) {
>>
>> Pass NULL as the first argument to __arch_prepare_bpf_trampoline()
>> call from arch_bpf_trampoline_size() function, to differentiate it
>> from how arch_prepare_bpf_trampoline() calls it and ensure maximum
>> possible instructions are emitted in arch_bpf_trampoline_size() for
>> ldimm instructions that load a different value during the actual JIT
>> compile of BPF trampoline.
>>
>> Fixes: d243b62b7bd3 ("powerpc64/bpf: Add support for bpf trampolines")
>> Reported-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
>> Closes: https://lore.kernel.org/all/6168bfc8-659f-4b5a-a6fb-90a916dde3b3@linux.ibm.com/
>> Cc: stable@vger.kernel.org # v6.13+
>> Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
>> ---
>>
>> * Removed a redundant '/' accidently added in a comment and resending.
>>
>>   arch/powerpc/net/bpf_jit_comp.c | 29 +++++++++++++++++++++++------
>>   1 file changed, 23 insertions(+), 6 deletions(-)
>>
>> diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
>> index 2991bb171a9b..c94717ccb2bd 100644
>> --- a/arch/powerpc/net/bpf_jit_comp.c
>> +++ b/arch/powerpc/net/bpf_jit_comp.c
>> @@ -833,7 +833,12 @@ static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_im
>>   	EMIT(PPC_RAW_STL(_R26, _R1, nvr_off + SZL));
>>   
>>   	if (flags & BPF_TRAMP_F_CALL_ORIG) {
>> -		PPC_LI_ADDR(_R3, (unsigned long)im);
>> +		/*
>> +		 * Emit maximum possible instructions while getting the size of
>> +		 * bpf trampoline to ensure trampoline JIT code doesn't overflow.
>> +		 */
>> +		PPC_LI_ADDR(_R3, im ? (unsigned long)im :
>> +				(unsigned long)(~(1UL << (BITS_PER_LONG - 1))));
> 
> We generally rely on  a NULL 'image' to detect a dummy pass. See commit
> d3921cbb6cd6 ("powerpc/bpf: Only pad length-variable code at initial
> pass"), for instance. Have you considered updating PPC_LI64() and
> PPC_LI32() to simply emit a fixed number of nops if image is NULL?

Did want to use image as NULL for the dummy pass but decided to do it
as a clean up later with bpf_jit_emit_func_call_rel(), PPC_LI64(),
PPC_LI32() and create_branch() needing it. But on second thoughts,
we are probably better off by accounting for max possible instructions
for all those cases (in the dummy pass) as part of this fix itself.
Will send a V2 soon...

Thanks
Hari